GET: Goal-directed Exploration and Targeting for Large-Scale Unknown Environments

Integrating LLM-based Semantic Reasoning and Memory for Autonomous Robot Search

Published

May 27, 2025

Authors: L. Zheng et al.
Published on Arxiv: 2025-05-27
Link: http://arxiv.org/abs/2505.20828v1
Institutions: Sun Yat-Sen University, School of Computer Science and Engineering • Sun Yat-Sen University, School of Systems Science and Engineering • Sun Yat-Sen University, School of Artificial Intelligence
Keywords: Large Language Models, LLM reasoning, Embodied AI, Object search, Goal-directed exploration, Diagram of Unified Thought (DoUT), Gaussian Mixture Model, Semantic octomap, Probabilistic memory, Autonomous robotics, Real-world experiments, Trajectory optimization, Spatial reasoning

Random Unsplash-style image

Searching for objects in vast, unstructured environments poses significant challenges for autonomous robots, especially under dynamic or outdoor conditions. Traditional approaches, often based on heuristics or hand-crafted routines, typically fail to generalize, while modern deep learning and reinforcement learning algorithms demand enormous data and may still lack robust transferability. Large Language Models (LLMs) introduce semantic reasoning, yet they struggle with spatial reasoning and effective use of memory in physically grounded settings.

To bridge these gaps, the article presents a novel framework, GET, which proposes the following key approaches and contributions:

Building on this integrated approach, the results highlight the significant performance gains achieved by GET:

Drawing on these comprehensive results, the study concludes with the following takeaways: